From 1fe996db6a98d4a6d3bc9f2427e57075015a05c7 Mon Sep 17 00:00:00 2001
From: Lek Huda <mhmmd.huda@gmail.com>
Date: Thu, 3 Sep 2026 09:12:20 +0700
Subject: [PATCH] fix(translator): route Gemini thinking through
 reasoning_effort on OpenAI-compatible wire

---
 .../translator/concerns/thinkingUnified.js    |  8 ++++++--
 tests/translator/thinking-unified.test.js     | 20 +++++++++++++++++++
 2 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/open-sse/translator/concerns/thinkingUnified.js b/open-sse/translator/concerns/thinkingUnified.js
index 5ea27b2e..a3035823 100644
--- a/open-sse/translator/concerns/thinkingUnified.js
+++ b/open-sse/translator/concerns/thinkingUnified.js
@@ -105,12 +105,16 @@ export function extractThinking(body) {
 // at the call-site where intent is snapshotted before format translation.
 export const captureThinking = extractThinking;
 
-// Resolve thinking format: provider override > capability > derive(targetFormat).
+const NATIVE_ONLY_FORMATS = new Set(["gemini-level", "gemini-budget", "claude-budget", "claude-adaptive", "kiro"]);
+
 function resolveFormat(targetFormat, model, provider) {
   const providerFmt = provider ? PROVIDERS[provider]?.thinkingFormat : null;
   if (providerFmt) return providerFmt;
   const caps = getCapabilitiesForModel(provider, model);
-  if (caps.thinkingFormat) return caps.thinkingFormat;
+  const isOpenAIWire = targetFormat === "openai" || targetFormat === "openai-responses";
+  if (caps.thinkingFormat && !(isOpenAIWire && NATIVE_ONLY_FORMATS.has(caps.thinkingFormat))) {
+    return caps.thinkingFormat;
+  }
   return FORMAT_TO_NATIVE[targetFormat] || "openai";
 }
 
diff --git a/tests/translator/thinking-unified.test.js b/tests/translator/thinking-unified.test.js
index 9ce75db5..93d8b683 100644
--- a/tests/translator/thinking-unified.test.js
+++ b/tests/translator/thinking-unified.test.js
@@ -220,6 +220,26 @@ describe("applyThinking per provider format", () => {
     const out = apply("openai", "gpt-5.6-sol", { reasoning_effort: "max" }, "kiro");
     expect(out.reasoning_effort).toBe("xhigh");
   });
+  it.each([
+    ["gemini-3.5-flash-lite"],
+    ["gemini-3.7-flash"],
+    ["gemini-3-pro"],
+  ])("Gemini 3.x model %s (gemini-level) over a custom OpenAI-compatible provider → reasoning_effort, not generationConfig (regression: #3718)", (model) => {
+    const out = apply("openai", model, { reasoning_effort: "medium" }, "my-custom-gemini-openai");
+    expect(out.reasoning_effort).toBe("medium");
+    expect(out.generationConfig).toBeUndefined();
+    expect(out.thinkingConfig).toBeUndefined();
+  });
+  it("Gemini 2.5 model (gemini-budget) over a custom OpenAI-compatible provider → reasoning_effort, not generationConfig (regression: #3718)", () => {
+    const out = apply("openai", "gemini-2.5-flash", { reasoning_effort: "high" }, "my-custom-gemini-openai");
+    expect(out.reasoning_effort).toBe("high");
+    expect(out.generationConfig).toBeUndefined();
+    expect(out.thinkingConfig).toBeUndefined();
+  });
+  it("Gemini model over its native format (antigravity/gemini-cli/vertex) still gets generationConfig", () => {
+    const out = apply("gemini-cli", "gemini-3.5-flash-lite", { reasoning_effort: "medium" }, "gemini-cli");
+    expect(out.generationConfig.thinkingConfig.thinkingLevel).toBe("medium");
+  });
 });
 
 describe("extractReasoningText (response shapes)", () => {